Reasoning about Exceptions during Plan Execution Monitoring

Carol A. Broverman, W. Bruce CroftAAAI 1987

Introduction

In the mid-1980s, AI planning began transitioning from closed-world simulations to open-world cooperative environments, such as office automation. In these settings, a hierarchical planner acts as a support tool rather than an omnipotent commander. Broverman and Croft (1987) identify a major friction point in this transition: Exceptions.

In traditional systems, if a user performs an action out of order or provides an unexpected type of data, the system views it as a failure. The standard response was "reactive" replanning—essentially trying to force the world back into the state the planner originally intended. The authors argue this is a "brainless" approach because it often ignores the user's deliberate intent, such as taking a shortcut. They propose a "constructive" approach where the system uses domain knowledge to reason about why an exception occurred, attempting to incorporate the user's unexpected actions into an evolving, consistent plan through negotiation.

System Architecture

The authors propose a general architecture that extends the traditional planning loop with two novel components: the Exception Analyst and the Negotiator.

The proposed system architecture, highlighting the roles of the Exception Analyst and Negotiator in the planning loop.

When the Execution Monitor detects a deviation from the "expected-action list," the Exception Analyst performs a controlled search through the knowledge base. It looks for common ancestors in the object hierarchy or "transformation" links to see if the user's action is semantically related to the expected one. If a relationship is found, the Negotiator facilitates a dialogue. It identifies "affected agents" (parts of the plan that might break) and "effecting agents" (the user) to reach a consensus on how to modify the plan.

Knowledge Representation and Reasoning

Central to this approach is a uniform Object-Based Representation for activities, objects, and agents. This includes a taxonomic hierarchy (is-a links) and "transformation" relationships (e.g., knowing that a phone number can be used to find an address).

The object-based representation hierarchy, showing how activities and objects are categorized to facilitate semantic reasoning.

By encoding how one object can be derived from another, the system can understand why a user might provide a "wrong" parameter that actually leads to the right result. Plans are maintained as procedural nets containing goal, action, and "phantom" nodes (goals already satisfied), linked by causal structures.

Categorizing Exceptions

The paper categorizes exceptions into four distinct classes that an interactive planner must handle:

  1. Step-out-of-order: The user performs a future step early (a shortcut).
  2. Unexpected-action: The user performs an action not in the plan at all.
  3. User-assertion: The user manually tells the system a goal is met (e.g., "I have the approval in my hand") without the system seeing the electronic trigger.
  4. Unexpected parameter type: The user provides a phone number where an address was expected.

Conclusion and Limitations

While this framework allows for "learning through failure"—where successful negotiation augments static domain plans—it remains heavily dependent on the quality of the knowledge base. If a transformation link is missing, the system cannot explain the exception. Additionally, at the time of publication, the implementation was a prototype, and the technical hurdles of propagating changes through an object-based representation without unintended side effects remained a significant challenge.

Summary updated

BibTeX
@inproceedings{broverman1987reasoning,
  author = {Carol A. Broverman and W. Bruce Croft},
  title = {{Reasoning about Exceptions during Plan Execution Monitoring}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure